Multimodal Causal Inference Analysis of Urban Landscape Therapeutic Effects with XGBoost and DiD

Abstract

Urban landscapes play a crucial role in promoting mental and physical health, yet existing methods often fail to integrate diverse data sources for assessing transformation impacts, leading to incomplete evaluations of therapeutic effects. This study addresses this gap by combining computer vision (CV) analysis of street view imagery, natural language processing (NLP) for text sentiment analysis, and environmental data (weather and air quality index) via XGBoost, followed by causal inference using difference-in-differences (DiD) to assess treatment effects of landscape changes. The aims are to develop a robust multimodal fusion model using XGBoost and apply difference-in-differences (DiD) for causal inference. Data from empirical datasets across 10 urban sites in cities like Beijing, changchun, and Nanjing (2020–2025) were analyzed, revealing a 45% increase in therapeutic indices post-transformation, with greenery visibility (35% importance) and positive sentiment (28%) as key predictors. Causal estimates confirmed significant health benefits, particularly in high-pollution areas (ATT=0.52, p<0.001). The findings demonstrate the efficacy of multimodal approaches over unimodal baselines (R²=0.85 vs. 0.58–0.72), highlighting green elements’ role in mitigating urban stressors. The study demonstrates the effectiveness of multimodal approaches in evaluating urban landscape transformations and their health benefits. The findings suggest that integrating green elements into urban planning can significantly enhance public health, especially in areas with high pollution. This framework can guide evidence-based urban planning for healthier cities.